1 | /* |
2 | * Copyright © 2008-2015 Intel Corporation |
3 | * |
4 | * Permission is hereby granted, free of charge, to any person obtaining a |
5 | * copy of this software and associated documentation files (the "Software"), |
6 | * to deal in the Software without restriction, including without limitation |
7 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, |
8 | * and/or sell copies of the Software, and to permit persons to whom the |
9 | * Software is furnished to do so, subject to the following conditions: |
10 | * |
11 | * The above copyright notice and this permission notice (including the next |
12 | * paragraph) shall be included in all copies or substantial portions of the |
13 | * Software. |
14 | * |
15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
16 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
17 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
18 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
19 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING |
20 | * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS |
21 | * IN THE SOFTWARE. |
22 | * |
23 | */ |
24 | |
25 | #include <linux/dma-fence-array.h> |
26 | #include <linux/dma-fence-chain.h> |
27 | #include <linux/irq_work.h> |
28 | #include <linux/prefetch.h> |
29 | #include <linux/sched.h> |
30 | #include <linux/sched/clock.h> |
31 | #include <linux/sched/signal.h> |
32 | #include <linux/sched/mm.h> |
33 | |
34 | #include "gem/i915_gem_context.h" |
35 | #include "gt/intel_breadcrumbs.h" |
36 | #include "gt/intel_context.h" |
37 | #include "gt/intel_engine.h" |
38 | #include "gt/intel_engine_heartbeat.h" |
39 | #include "gt/intel_engine_regs.h" |
40 | #include "gt/intel_gpu_commands.h" |
41 | #include "gt/intel_reset.h" |
42 | #include "gt/intel_ring.h" |
43 | #include "gt/intel_rps.h" |
44 | |
45 | #include "i915_active.h" |
46 | #include "i915_config.h" |
47 | #include "i915_deps.h" |
48 | #include "i915_driver.h" |
49 | #include "i915_drv.h" |
50 | #include "i915_trace.h" |
51 | |
52 | struct execute_cb { |
53 | struct irq_work work; |
54 | struct i915_sw_fence *fence; |
55 | struct i915_request *signal; |
56 | }; |
57 | |
58 | static struct kmem_cache *slab_requests; |
59 | static struct kmem_cache *slab_execute_cbs; |
60 | |
61 | static const char *i915_fence_get_driver_name(struct dma_fence *fence) |
62 | { |
63 | return dev_name(dev: to_request(fence)->i915->drm.dev); |
64 | } |
65 | |
66 | static const char *i915_fence_get_timeline_name(struct dma_fence *fence) |
67 | { |
68 | const struct i915_gem_context *ctx; |
69 | |
70 | /* |
71 | * The timeline struct (as part of the ppgtt underneath a context) |
72 | * may be freed when the request is no longer in use by the GPU. |
73 | * We could extend the life of a context to beyond that of all |
74 | * fences, possibly keeping the hw resource around indefinitely, |
75 | * or we just give them a false name. Since |
76 | * dma_fence_ops.get_timeline_name is a debug feature, the occasional |
77 | * lie seems justifiable. |
78 | */ |
79 | if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) |
80 | return "signaled" ; |
81 | |
82 | ctx = i915_request_gem_context(rq: to_request(fence)); |
83 | if (!ctx) |
84 | return "[" DRIVER_NAME "]" ; |
85 | |
86 | return ctx->name; |
87 | } |
88 | |
89 | static bool i915_fence_signaled(struct dma_fence *fence) |
90 | { |
91 | return i915_request_completed(rq: to_request(fence)); |
92 | } |
93 | |
94 | static bool i915_fence_enable_signaling(struct dma_fence *fence) |
95 | { |
96 | return i915_request_enable_breadcrumb(request: to_request(fence)); |
97 | } |
98 | |
99 | static signed long i915_fence_wait(struct dma_fence *fence, |
100 | bool interruptible, |
101 | signed long timeout) |
102 | { |
103 | return i915_request_wait_timeout(rq: to_request(fence), |
104 | flags: interruptible | I915_WAIT_PRIORITY, |
105 | timeout); |
106 | } |
107 | |
108 | struct kmem_cache *i915_request_slab_cache(void) |
109 | { |
110 | return slab_requests; |
111 | } |
112 | |
113 | static void i915_fence_release(struct dma_fence *fence) |
114 | { |
115 | struct i915_request *rq = to_request(fence); |
116 | |
117 | GEM_BUG_ON(rq->guc_prio != GUC_PRIO_INIT && |
118 | rq->guc_prio != GUC_PRIO_FINI); |
119 | |
120 | i915_request_free_capture_list(fetch_and_zero(&rq->capture_list)); |
121 | if (rq->batch_res) { |
122 | i915_vma_resource_put(vma_res: rq->batch_res); |
123 | rq->batch_res = NULL; |
124 | } |
125 | |
126 | /* |
127 | * The request is put onto a RCU freelist (i.e. the address |
128 | * is immediately reused), mark the fences as being freed now. |
129 | * Otherwise the debugobjects for the fences are only marked as |
130 | * freed when the slab cache itself is freed, and so we would get |
131 | * caught trying to reuse dead objects. |
132 | */ |
133 | i915_sw_fence_fini(fence: &rq->submit); |
134 | i915_sw_fence_fini(fence: &rq->semaphore); |
135 | |
136 | /* |
137 | * Keep one request on each engine for reserved use under mempressure. |
138 | * |
139 | * We do not hold a reference to the engine here and so have to be |
140 | * very careful in what rq->engine we poke. The virtual engine is |
141 | * referenced via the rq->context and we released that ref during |
142 | * i915_request_retire(), ergo we must not dereference a virtual |
143 | * engine here. Not that we would want to, as the only consumer of |
144 | * the reserved engine->request_pool is the power management parking, |
145 | * which must-not-fail, and that is only run on the physical engines. |
146 | * |
147 | * Since the request must have been executed to be have completed, |
148 | * we know that it will have been processed by the HW and will |
149 | * not be unsubmitted again, so rq->engine and rq->execution_mask |
150 | * at this point is stable. rq->execution_mask will be a single |
151 | * bit if the last and _only_ engine it could execution on was a |
152 | * physical engine, if it's multiple bits then it started on and |
153 | * could still be on a virtual engine. Thus if the mask is not a |
154 | * power-of-two we assume that rq->engine may still be a virtual |
155 | * engine and so a dangling invalid pointer that we cannot dereference |
156 | * |
157 | * For example, consider the flow of a bonded request through a virtual |
158 | * engine. The request is created with a wide engine mask (all engines |
159 | * that we might execute on). On processing the bond, the request mask |
160 | * is reduced to one or more engines. If the request is subsequently |
161 | * bound to a single engine, it will then be constrained to only |
162 | * execute on that engine and never returned to the virtual engine |
163 | * after timeslicing away, see __unwind_incomplete_requests(). Thus we |
164 | * know that if the rq->execution_mask is a single bit, rq->engine |
165 | * can be a physical engine with the exact corresponding mask. |
166 | */ |
167 | if (is_power_of_2(n: rq->execution_mask) && |
168 | !cmpxchg(&rq->engine->request_pool, NULL, rq)) |
169 | return; |
170 | |
171 | kmem_cache_free(s: slab_requests, objp: rq); |
172 | } |
173 | |
174 | const struct dma_fence_ops i915_fence_ops = { |
175 | .get_driver_name = i915_fence_get_driver_name, |
176 | .get_timeline_name = i915_fence_get_timeline_name, |
177 | .enable_signaling = i915_fence_enable_signaling, |
178 | .signaled = i915_fence_signaled, |
179 | .wait = i915_fence_wait, |
180 | .release = i915_fence_release, |
181 | }; |
182 | |
183 | static void irq_execute_cb(struct irq_work *wrk) |
184 | { |
185 | struct execute_cb *cb = container_of(wrk, typeof(*cb), work); |
186 | |
187 | i915_sw_fence_complete(fence: cb->fence); |
188 | kmem_cache_free(s: slab_execute_cbs, objp: cb); |
189 | } |
190 | |
191 | static __always_inline void |
192 | __notify_execute_cb(struct i915_request *rq, bool (*fn)(struct irq_work *wrk)) |
193 | { |
194 | struct execute_cb *cb, *cn; |
195 | |
196 | if (llist_empty(head: &rq->execute_cb)) |
197 | return; |
198 | |
199 | llist_for_each_entry_safe(cb, cn, |
200 | llist_del_all(&rq->execute_cb), |
201 | work.node.llist) |
202 | fn(&cb->work); |
203 | } |
204 | |
205 | static void __notify_execute_cb_irq(struct i915_request *rq) |
206 | { |
207 | __notify_execute_cb(rq, fn: irq_work_queue); |
208 | } |
209 | |
210 | static bool irq_work_imm(struct irq_work *wrk) |
211 | { |
212 | wrk->func(wrk); |
213 | return false; |
214 | } |
215 | |
216 | void i915_request_notify_execute_cb_imm(struct i915_request *rq) |
217 | { |
218 | __notify_execute_cb(rq, fn: irq_work_imm); |
219 | } |
220 | |
221 | static void __i915_request_fill(struct i915_request *rq, u8 val) |
222 | { |
223 | void *vaddr = rq->ring->vaddr; |
224 | u32 head; |
225 | |
226 | head = rq->infix; |
227 | if (rq->postfix < head) { |
228 | memset(vaddr + head, val, rq->ring->size - head); |
229 | head = 0; |
230 | } |
231 | memset(vaddr + head, val, rq->postfix - head); |
232 | } |
233 | |
234 | /** |
235 | * i915_request_active_engine |
236 | * @rq: request to inspect |
237 | * @active: pointer in which to return the active engine |
238 | * |
239 | * Fills the currently active engine to the @active pointer if the request |
240 | * is active and still not completed. |
241 | * |
242 | * Returns true if request was active or false otherwise. |
243 | */ |
244 | bool |
245 | i915_request_active_engine(struct i915_request *rq, |
246 | struct intel_engine_cs **active) |
247 | { |
248 | struct intel_engine_cs *engine, *locked; |
249 | bool ret = false; |
250 | |
251 | /* |
252 | * Serialise with __i915_request_submit() so that it sees |
253 | * is-banned?, or we know the request is already inflight. |
254 | * |
255 | * Note that rq->engine is unstable, and so we double |
256 | * check that we have acquired the lock on the final engine. |
257 | */ |
258 | locked = READ_ONCE(rq->engine); |
259 | spin_lock_irq(lock: &locked->sched_engine->lock); |
260 | while (unlikely(locked != (engine = READ_ONCE(rq->engine)))) { |
261 | spin_unlock(lock: &locked->sched_engine->lock); |
262 | locked = engine; |
263 | spin_lock(lock: &locked->sched_engine->lock); |
264 | } |
265 | |
266 | if (i915_request_is_active(rq)) { |
267 | if (!__i915_request_is_complete(rq)) |
268 | *active = locked; |
269 | ret = true; |
270 | } |
271 | |
272 | spin_unlock_irq(lock: &locked->sched_engine->lock); |
273 | |
274 | return ret; |
275 | } |
276 | |
277 | static void __rq_init_watchdog(struct i915_request *rq) |
278 | { |
279 | rq->watchdog.timer.function = NULL; |
280 | } |
281 | |
282 | static enum hrtimer_restart __rq_watchdog_expired(struct hrtimer *hrtimer) |
283 | { |
284 | struct i915_request *rq = |
285 | container_of(hrtimer, struct i915_request, watchdog.timer); |
286 | struct intel_gt *gt = rq->engine->gt; |
287 | |
288 | if (!i915_request_completed(rq)) { |
289 | if (llist_add(new: &rq->watchdog.link, head: >->watchdog.list)) |
290 | queue_work(wq: gt->i915->unordered_wq, work: >->watchdog.work); |
291 | } else { |
292 | i915_request_put(rq); |
293 | } |
294 | |
295 | return HRTIMER_NORESTART; |
296 | } |
297 | |
298 | static void __rq_arm_watchdog(struct i915_request *rq) |
299 | { |
300 | struct i915_request_watchdog *wdg = &rq->watchdog; |
301 | struct intel_context *ce = rq->context; |
302 | |
303 | if (!ce->watchdog.timeout_us) |
304 | return; |
305 | |
306 | i915_request_get(rq); |
307 | |
308 | hrtimer_init(timer: &wdg->timer, CLOCK_MONOTONIC, mode: HRTIMER_MODE_REL); |
309 | wdg->timer.function = __rq_watchdog_expired; |
310 | hrtimer_start_range_ns(timer: &wdg->timer, |
311 | tim: ns_to_ktime(ns: ce->watchdog.timeout_us * |
312 | NSEC_PER_USEC), |
313 | NSEC_PER_MSEC, |
314 | mode: HRTIMER_MODE_REL); |
315 | } |
316 | |
317 | static void __rq_cancel_watchdog(struct i915_request *rq) |
318 | { |
319 | struct i915_request_watchdog *wdg = &rq->watchdog; |
320 | |
321 | if (wdg->timer.function && hrtimer_try_to_cancel(timer: &wdg->timer) > 0) |
322 | i915_request_put(rq); |
323 | } |
324 | |
325 | #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR) |
326 | |
327 | /** |
328 | * i915_request_free_capture_list - Free a capture list |
329 | * @capture: Pointer to the first list item or NULL |
330 | * |
331 | */ |
332 | void i915_request_free_capture_list(struct i915_capture_list *capture) |
333 | { |
334 | while (capture) { |
335 | struct i915_capture_list *next = capture->next; |
336 | |
337 | i915_vma_resource_put(vma_res: capture->vma_res); |
338 | kfree(objp: capture); |
339 | capture = next; |
340 | } |
341 | } |
342 | |
343 | #define assert_capture_list_is_null(_rq) GEM_BUG_ON((_rq)->capture_list) |
344 | |
345 | #define clear_capture_list(_rq) ((_rq)->capture_list = NULL) |
346 | |
347 | #else |
348 | |
349 | #define i915_request_free_capture_list(_a) do {} while (0) |
350 | |
351 | #define assert_capture_list_is_null(_a) do {} while (0) |
352 | |
353 | #define clear_capture_list(_rq) do {} while (0) |
354 | |
355 | #endif |
356 | |
357 | bool i915_request_retire(struct i915_request *rq) |
358 | { |
359 | if (!__i915_request_is_complete(rq)) |
360 | return false; |
361 | |
362 | RQ_TRACE(rq, "\n" ); |
363 | |
364 | GEM_BUG_ON(!i915_sw_fence_signaled(&rq->submit)); |
365 | trace_i915_request_retire(rq); |
366 | i915_request_mark_complete(rq); |
367 | |
368 | __rq_cancel_watchdog(rq); |
369 | |
370 | /* |
371 | * We know the GPU must have read the request to have |
372 | * sent us the seqno + interrupt, so use the position |
373 | * of tail of the request to update the last known position |
374 | * of the GPU head. |
375 | * |
376 | * Note this requires that we are always called in request |
377 | * completion order. |
378 | */ |
379 | GEM_BUG_ON(!list_is_first(&rq->link, |
380 | &i915_request_timeline(rq)->requests)); |
381 | if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)) |
382 | /* Poison before we release our space in the ring */ |
383 | __i915_request_fill(rq, POISON_FREE); |
384 | rq->ring->head = rq->postfix; |
385 | |
386 | if (!i915_request_signaled(rq)) { |
387 | spin_lock_irq(lock: &rq->lock); |
388 | dma_fence_signal_locked(fence: &rq->fence); |
389 | spin_unlock_irq(lock: &rq->lock); |
390 | } |
391 | |
392 | if (test_and_set_bit(nr: I915_FENCE_FLAG_BOOST, addr: &rq->fence.flags)) |
393 | intel_rps_dec_waiters(rps: &rq->engine->gt->rps); |
394 | |
395 | /* |
396 | * We only loosely track inflight requests across preemption, |
397 | * and so we may find ourselves attempting to retire a _completed_ |
398 | * request that we have removed from the HW and put back on a run |
399 | * queue. |
400 | * |
401 | * As we set I915_FENCE_FLAG_ACTIVE on the request, this should be |
402 | * after removing the breadcrumb and signaling it, so that we do not |
403 | * inadvertently attach the breadcrumb to a completed request. |
404 | */ |
405 | rq->engine->remove_active_request(rq); |
406 | GEM_BUG_ON(!llist_empty(&rq->execute_cb)); |
407 | |
408 | __list_del_entry(entry: &rq->link); /* poison neither prev/next (RCU walks) */ |
409 | |
410 | intel_context_exit(ce: rq->context); |
411 | intel_context_unpin(ce: rq->context); |
412 | |
413 | i915_sched_node_fini(node: &rq->sched); |
414 | i915_request_put(rq); |
415 | |
416 | return true; |
417 | } |
418 | |
419 | void i915_request_retire_upto(struct i915_request *rq) |
420 | { |
421 | struct intel_timeline * const tl = i915_request_timeline(rq); |
422 | struct i915_request *tmp; |
423 | |
424 | RQ_TRACE(rq, "\n" ); |
425 | GEM_BUG_ON(!__i915_request_is_complete(rq)); |
426 | |
427 | do { |
428 | tmp = list_first_entry(&tl->requests, typeof(*tmp), link); |
429 | GEM_BUG_ON(!i915_request_completed(tmp)); |
430 | } while (i915_request_retire(rq: tmp) && tmp != rq); |
431 | } |
432 | |
433 | static struct i915_request * const * |
434 | __engine_active(struct intel_engine_cs *engine) |
435 | { |
436 | return READ_ONCE(engine->execlists.active); |
437 | } |
438 | |
439 | static bool __request_in_flight(const struct i915_request *signal) |
440 | { |
441 | struct i915_request * const *port, *rq; |
442 | bool inflight = false; |
443 | |
444 | if (!i915_request_is_ready(rq: signal)) |
445 | return false; |
446 | |
447 | /* |
448 | * Even if we have unwound the request, it may still be on |
449 | * the GPU (preempt-to-busy). If that request is inside an |
450 | * unpreemptible critical section, it will not be removed. Some |
451 | * GPU functions may even be stuck waiting for the paired request |
452 | * (__await_execution) to be submitted and cannot be preempted |
453 | * until the bond is executing. |
454 | * |
455 | * As we know that there are always preemption points between |
456 | * requests, we know that only the currently executing request |
457 | * may be still active even though we have cleared the flag. |
458 | * However, we can't rely on our tracking of ELSP[0] to know |
459 | * which request is currently active and so maybe stuck, as |
460 | * the tracking maybe an event behind. Instead assume that |
461 | * if the context is still inflight, then it is still active |
462 | * even if the active flag has been cleared. |
463 | * |
464 | * To further complicate matters, if there a pending promotion, the HW |
465 | * may either perform a context switch to the second inflight execlists, |
466 | * or it may switch to the pending set of execlists. In the case of the |
467 | * latter, it may send the ACK and we process the event copying the |
468 | * pending[] over top of inflight[], _overwriting_ our *active. Since |
469 | * this implies the HW is arbitrating and not struck in *active, we do |
470 | * not worry about complete accuracy, but we do require no read/write |
471 | * tearing of the pointer [the read of the pointer must be valid, even |
472 | * as the array is being overwritten, for which we require the writes |
473 | * to avoid tearing.] |
474 | * |
475 | * Note that the read of *execlists->active may race with the promotion |
476 | * of execlists->pending[] to execlists->inflight[], overwritting |
477 | * the value at *execlists->active. This is fine. The promotion implies |
478 | * that we received an ACK from the HW, and so the context is not |
479 | * stuck -- if we do not see ourselves in *active, the inflight status |
480 | * is valid. If instead we see ourselves being copied into *active, |
481 | * we are inflight and may signal the callback. |
482 | */ |
483 | if (!intel_context_inflight(signal->context)) |
484 | return false; |
485 | |
486 | rcu_read_lock(); |
487 | for (port = __engine_active(engine: signal->engine); |
488 | (rq = READ_ONCE(*port)); /* may race with promotion of pending[] */ |
489 | port++) { |
490 | if (rq->context == signal->context) { |
491 | inflight = i915_seqno_passed(seq1: rq->fence.seqno, |
492 | seq2: signal->fence.seqno); |
493 | break; |
494 | } |
495 | } |
496 | rcu_read_unlock(); |
497 | |
498 | return inflight; |
499 | } |
500 | |
501 | static int |
502 | __await_execution(struct i915_request *rq, |
503 | struct i915_request *signal, |
504 | gfp_t gfp) |
505 | { |
506 | struct execute_cb *cb; |
507 | |
508 | if (i915_request_is_active(rq: signal)) |
509 | return 0; |
510 | |
511 | cb = kmem_cache_alloc(cachep: slab_execute_cbs, flags: gfp); |
512 | if (!cb) |
513 | return -ENOMEM; |
514 | |
515 | cb->fence = &rq->submit; |
516 | i915_sw_fence_await(fence: cb->fence); |
517 | init_irq_work(work: &cb->work, func: irq_execute_cb); |
518 | |
519 | /* |
520 | * Register the callback first, then see if the signaler is already |
521 | * active. This ensures that if we race with the |
522 | * __notify_execute_cb from i915_request_submit() and we are not |
523 | * included in that list, we get a second bite of the cherry and |
524 | * execute it ourselves. After this point, a future |
525 | * i915_request_submit() will notify us. |
526 | * |
527 | * In i915_request_retire() we set the ACTIVE bit on a completed |
528 | * request (then flush the execute_cb). So by registering the |
529 | * callback first, then checking the ACTIVE bit, we serialise with |
530 | * the completed/retired request. |
531 | */ |
532 | if (llist_add(new: &cb->work.node.llist, head: &signal->execute_cb)) { |
533 | if (i915_request_is_active(rq: signal) || |
534 | __request_in_flight(signal)) |
535 | i915_request_notify_execute_cb_imm(rq: signal); |
536 | } |
537 | |
538 | return 0; |
539 | } |
540 | |
541 | static bool fatal_error(int error) |
542 | { |
543 | switch (error) { |
544 | case 0: /* not an error! */ |
545 | case -EAGAIN: /* innocent victim of a GT reset (__i915_request_reset) */ |
546 | case -ETIMEDOUT: /* waiting for Godot (timer_i915_sw_fence_wake) */ |
547 | return false; |
548 | default: |
549 | return true; |
550 | } |
551 | } |
552 | |
553 | void __i915_request_skip(struct i915_request *rq) |
554 | { |
555 | GEM_BUG_ON(!fatal_error(rq->fence.error)); |
556 | |
557 | if (rq->infix == rq->postfix) |
558 | return; |
559 | |
560 | RQ_TRACE(rq, "error: %d\n" , rq->fence.error); |
561 | |
562 | /* |
563 | * As this request likely depends on state from the lost |
564 | * context, clear out all the user operations leaving the |
565 | * breadcrumb at the end (so we get the fence notifications). |
566 | */ |
567 | __i915_request_fill(rq, val: 0); |
568 | rq->infix = rq->postfix; |
569 | } |
570 | |
571 | bool i915_request_set_error_once(struct i915_request *rq, int error) |
572 | { |
573 | int old; |
574 | |
575 | GEM_BUG_ON(!IS_ERR_VALUE((long)error)); |
576 | |
577 | if (i915_request_signaled(rq)) |
578 | return false; |
579 | |
580 | old = READ_ONCE(rq->fence.error); |
581 | do { |
582 | if (fatal_error(error: old)) |
583 | return false; |
584 | } while (!try_cmpxchg(&rq->fence.error, &old, error)); |
585 | |
586 | return true; |
587 | } |
588 | |
589 | struct i915_request *i915_request_mark_eio(struct i915_request *rq) |
590 | { |
591 | if (__i915_request_is_complete(rq)) |
592 | return NULL; |
593 | |
594 | GEM_BUG_ON(i915_request_signaled(rq)); |
595 | |
596 | /* As soon as the request is completed, it may be retired */ |
597 | rq = i915_request_get(rq); |
598 | |
599 | i915_request_set_error_once(rq, error: -EIO); |
600 | i915_request_mark_complete(rq); |
601 | |
602 | return rq; |
603 | } |
604 | |
605 | bool __i915_request_submit(struct i915_request *request) |
606 | { |
607 | struct intel_engine_cs *engine = request->engine; |
608 | bool result = false; |
609 | |
610 | RQ_TRACE(request, "\n" ); |
611 | |
612 | GEM_BUG_ON(!irqs_disabled()); |
613 | lockdep_assert_held(&engine->sched_engine->lock); |
614 | |
615 | /* |
616 | * With the advent of preempt-to-busy, we frequently encounter |
617 | * requests that we have unsubmitted from HW, but left running |
618 | * until the next ack and so have completed in the meantime. On |
619 | * resubmission of that completed request, we can skip |
620 | * updating the payload, and execlists can even skip submitting |
621 | * the request. |
622 | * |
623 | * We must remove the request from the caller's priority queue, |
624 | * and the caller must only call us when the request is in their |
625 | * priority queue, under the sched_engine->lock. This ensures that the |
626 | * request has *not* yet been retired and we can safely move |
627 | * the request into the engine->active.list where it will be |
628 | * dropped upon retiring. (Otherwise if resubmit a *retired* |
629 | * request, this would be a horrible use-after-free.) |
630 | */ |
631 | if (__i915_request_is_complete(rq: request)) { |
632 | list_del_init(entry: &request->sched.link); |
633 | goto active; |
634 | } |
635 | |
636 | if (unlikely(!intel_context_is_schedulable(request->context))) |
637 | i915_request_set_error_once(rq: request, error: -EIO); |
638 | |
639 | if (unlikely(fatal_error(request->fence.error))) |
640 | __i915_request_skip(rq: request); |
641 | |
642 | /* |
643 | * Are we using semaphores when the gpu is already saturated? |
644 | * |
645 | * Using semaphores incurs a cost in having the GPU poll a |
646 | * memory location, busywaiting for it to change. The continual |
647 | * memory reads can have a noticeable impact on the rest of the |
648 | * system with the extra bus traffic, stalling the cpu as it too |
649 | * tries to access memory across the bus (perf stat -e bus-cycles). |
650 | * |
651 | * If we installed a semaphore on this request and we only submit |
652 | * the request after the signaler completed, that indicates the |
653 | * system is overloaded and using semaphores at this time only |
654 | * increases the amount of work we are doing. If so, we disable |
655 | * further use of semaphores until we are idle again, whence we |
656 | * optimistically try again. |
657 | */ |
658 | if (request->sched.semaphores && |
659 | i915_sw_fence_signaled(fence: &request->semaphore)) |
660 | engine->saturated |= request->sched.semaphores; |
661 | |
662 | engine->emit_fini_breadcrumb(request, |
663 | request->ring->vaddr + request->postfix); |
664 | |
665 | trace_i915_request_execute(rq: request); |
666 | if (engine->bump_serial) |
667 | engine->bump_serial(engine); |
668 | else |
669 | engine->serial++; |
670 | |
671 | result = true; |
672 | |
673 | GEM_BUG_ON(test_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags)); |
674 | engine->add_active_request(request); |
675 | active: |
676 | clear_bit(nr: I915_FENCE_FLAG_PQUEUE, addr: &request->fence.flags); |
677 | set_bit(nr: I915_FENCE_FLAG_ACTIVE, addr: &request->fence.flags); |
678 | |
679 | /* |
680 | * XXX Rollback bonded-execution on __i915_request_unsubmit()? |
681 | * |
682 | * In the future, perhaps when we have an active time-slicing scheduler, |
683 | * it will be interesting to unsubmit parallel execution and remove |
684 | * busywaits from the GPU until their master is restarted. This is |
685 | * quite hairy, we have to carefully rollback the fence and do a |
686 | * preempt-to-idle cycle on the target engine, all the while the |
687 | * master execute_cb may refire. |
688 | */ |
689 | __notify_execute_cb_irq(rq: request); |
690 | |
691 | /* We may be recursing from the signal callback of another i915 fence */ |
692 | if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags)) |
693 | i915_request_enable_breadcrumb(request); |
694 | |
695 | return result; |
696 | } |
697 | |
698 | void i915_request_submit(struct i915_request *request) |
699 | { |
700 | struct intel_engine_cs *engine = request->engine; |
701 | unsigned long flags; |
702 | |
703 | /* Will be called from irq-context when using foreign fences. */ |
704 | spin_lock_irqsave(&engine->sched_engine->lock, flags); |
705 | |
706 | __i915_request_submit(request); |
707 | |
708 | spin_unlock_irqrestore(lock: &engine->sched_engine->lock, flags); |
709 | } |
710 | |
711 | void __i915_request_unsubmit(struct i915_request *request) |
712 | { |
713 | struct intel_engine_cs *engine = request->engine; |
714 | |
715 | /* |
716 | * Only unwind in reverse order, required so that the per-context list |
717 | * is kept in seqno/ring order. |
718 | */ |
719 | RQ_TRACE(request, "\n" ); |
720 | |
721 | GEM_BUG_ON(!irqs_disabled()); |
722 | lockdep_assert_held(&engine->sched_engine->lock); |
723 | |
724 | /* |
725 | * Before we remove this breadcrumb from the signal list, we have |
726 | * to ensure that a concurrent dma_fence_enable_signaling() does not |
727 | * attach itself. We first mark the request as no longer active and |
728 | * make sure that is visible to other cores, and then remove the |
729 | * breadcrumb if attached. |
730 | */ |
731 | GEM_BUG_ON(!test_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags)); |
732 | clear_bit_unlock(nr: I915_FENCE_FLAG_ACTIVE, addr: &request->fence.flags); |
733 | if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags)) |
734 | i915_request_cancel_breadcrumb(request); |
735 | |
736 | /* We've already spun, don't charge on resubmitting. */ |
737 | if (request->sched.semaphores && __i915_request_has_started(rq: request)) |
738 | request->sched.semaphores = 0; |
739 | |
740 | /* |
741 | * We don't need to wake_up any waiters on request->execute, they |
742 | * will get woken by any other event or us re-adding this request |
743 | * to the engine timeline (__i915_request_submit()). The waiters |
744 | * should be quite adapt at finding that the request now has a new |
745 | * global_seqno to the one they went to sleep on. |
746 | */ |
747 | } |
748 | |
749 | void i915_request_unsubmit(struct i915_request *request) |
750 | { |
751 | struct intel_engine_cs *engine = request->engine; |
752 | unsigned long flags; |
753 | |
754 | /* Will be called from irq-context when using foreign fences. */ |
755 | spin_lock_irqsave(&engine->sched_engine->lock, flags); |
756 | |
757 | __i915_request_unsubmit(request); |
758 | |
759 | spin_unlock_irqrestore(lock: &engine->sched_engine->lock, flags); |
760 | } |
761 | |
762 | void i915_request_cancel(struct i915_request *rq, int error) |
763 | { |
764 | if (!i915_request_set_error_once(rq, error)) |
765 | return; |
766 | |
767 | set_bit(nr: I915_FENCE_FLAG_SENTINEL, addr: &rq->fence.flags); |
768 | |
769 | intel_context_cancel_request(ce: rq->context, rq); |
770 | } |
771 | |
772 | static int |
773 | submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state) |
774 | { |
775 | struct i915_request *request = |
776 | container_of(fence, typeof(*request), submit); |
777 | |
778 | switch (state) { |
779 | case FENCE_COMPLETE: |
780 | trace_i915_request_submit(rq: request); |
781 | |
782 | if (unlikely(fence->error)) |
783 | i915_request_set_error_once(rq: request, error: fence->error); |
784 | else |
785 | __rq_arm_watchdog(rq: request); |
786 | |
787 | /* |
788 | * We need to serialize use of the submit_request() callback |
789 | * with its hotplugging performed during an emergency |
790 | * i915_gem_set_wedged(). We use the RCU mechanism to mark the |
791 | * critical section in order to force i915_gem_set_wedged() to |
792 | * wait until the submit_request() is completed before |
793 | * proceeding. |
794 | */ |
795 | rcu_read_lock(); |
796 | request->engine->submit_request(request); |
797 | rcu_read_unlock(); |
798 | break; |
799 | |
800 | case FENCE_FREE: |
801 | i915_request_put(rq: request); |
802 | break; |
803 | } |
804 | |
805 | return NOTIFY_DONE; |
806 | } |
807 | |
808 | static int |
809 | semaphore_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state) |
810 | { |
811 | struct i915_request *rq = container_of(fence, typeof(*rq), semaphore); |
812 | |
813 | switch (state) { |
814 | case FENCE_COMPLETE: |
815 | break; |
816 | |
817 | case FENCE_FREE: |
818 | i915_request_put(rq); |
819 | break; |
820 | } |
821 | |
822 | return NOTIFY_DONE; |
823 | } |
824 | |
825 | static void retire_requests(struct intel_timeline *tl) |
826 | { |
827 | struct i915_request *rq, *rn; |
828 | |
829 | list_for_each_entry_safe(rq, rn, &tl->requests, link) |
830 | if (!i915_request_retire(rq)) |
831 | break; |
832 | } |
833 | |
834 | static noinline struct i915_request * |
835 | request_alloc_slow(struct intel_timeline *tl, |
836 | struct i915_request **rsvd, |
837 | gfp_t gfp) |
838 | { |
839 | struct i915_request *rq; |
840 | |
841 | /* If we cannot wait, dip into our reserves */ |
842 | if (!gfpflags_allow_blocking(gfp_flags: gfp)) { |
843 | rq = xchg(rsvd, NULL); |
844 | if (!rq) /* Use the normal failure path for one final WARN */ |
845 | goto out; |
846 | |
847 | return rq; |
848 | } |
849 | |
850 | if (list_empty(head: &tl->requests)) |
851 | goto out; |
852 | |
853 | /* Move our oldest request to the slab-cache (if not in use!) */ |
854 | rq = list_first_entry(&tl->requests, typeof(*rq), link); |
855 | i915_request_retire(rq); |
856 | |
857 | rq = kmem_cache_alloc(cachep: slab_requests, |
858 | flags: gfp | __GFP_RETRY_MAYFAIL | __GFP_NOWARN); |
859 | if (rq) |
860 | return rq; |
861 | |
862 | /* Ratelimit ourselves to prevent oom from malicious clients */ |
863 | rq = list_last_entry(&tl->requests, typeof(*rq), link); |
864 | cond_synchronize_rcu(oldstate: rq->rcustate); |
865 | |
866 | /* Retire our old requests in the hope that we free some */ |
867 | retire_requests(tl); |
868 | |
869 | out: |
870 | return kmem_cache_alloc(cachep: slab_requests, flags: gfp); |
871 | } |
872 | |
873 | static void __i915_request_ctor(void *arg) |
874 | { |
875 | struct i915_request *rq = arg; |
876 | |
877 | spin_lock_init(&rq->lock); |
878 | i915_sched_node_init(node: &rq->sched); |
879 | i915_sw_fence_init(&rq->submit, submit_notify); |
880 | i915_sw_fence_init(&rq->semaphore, semaphore_notify); |
881 | |
882 | clear_capture_list(rq); |
883 | rq->batch_res = NULL; |
884 | |
885 | init_llist_head(list: &rq->execute_cb); |
886 | } |
887 | |
888 | #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) |
889 | #define clear_batch_ptr(_rq) ((_rq)->batch = NULL) |
890 | #else |
891 | #define clear_batch_ptr(_a) do {} while (0) |
892 | #endif |
893 | |
894 | struct i915_request * |
895 | __i915_request_create(struct intel_context *ce, gfp_t gfp) |
896 | { |
897 | struct intel_timeline *tl = ce->timeline; |
898 | struct i915_request *rq; |
899 | u32 seqno; |
900 | int ret; |
901 | |
902 | might_alloc(gfp_mask: gfp); |
903 | |
904 | /* Check that the caller provided an already pinned context */ |
905 | __intel_context_pin(ce); |
906 | |
907 | /* |
908 | * Beware: Dragons be flying overhead. |
909 | * |
910 | * We use RCU to look up requests in flight. The lookups may |
911 | * race with the request being allocated from the slab freelist. |
912 | * That is the request we are writing to here, may be in the process |
913 | * of being read by __i915_active_request_get_rcu(). As such, |
914 | * we have to be very careful when overwriting the contents. During |
915 | * the RCU lookup, we change chase the request->engine pointer, |
916 | * read the request->global_seqno and increment the reference count. |
917 | * |
918 | * The reference count is incremented atomically. If it is zero, |
919 | * the lookup knows the request is unallocated and complete. Otherwise, |
920 | * it is either still in use, or has been reallocated and reset |
921 | * with dma_fence_init(). This increment is safe for release as we |
922 | * check that the request we have a reference to and matches the active |
923 | * request. |
924 | * |
925 | * Before we increment the refcount, we chase the request->engine |
926 | * pointer. We must not call kmem_cache_zalloc() or else we set |
927 | * that pointer to NULL and cause a crash during the lookup. If |
928 | * we see the request is completed (based on the value of the |
929 | * old engine and seqno), the lookup is complete and reports NULL. |
930 | * If we decide the request is not completed (new engine or seqno), |
931 | * then we grab a reference and double check that it is still the |
932 | * active request - which it won't be and restart the lookup. |
933 | * |
934 | * Do not use kmem_cache_zalloc() here! |
935 | */ |
936 | rq = kmem_cache_alloc(cachep: slab_requests, |
937 | flags: gfp | __GFP_RETRY_MAYFAIL | __GFP_NOWARN); |
938 | if (unlikely(!rq)) { |
939 | rq = request_alloc_slow(tl, rsvd: &ce->engine->request_pool, gfp); |
940 | if (!rq) { |
941 | ret = -ENOMEM; |
942 | goto err_unreserve; |
943 | } |
944 | } |
945 | |
946 | rq->context = ce; |
947 | rq->engine = ce->engine; |
948 | rq->ring = ce->ring; |
949 | rq->execution_mask = ce->engine->mask; |
950 | rq->i915 = ce->engine->i915; |
951 | |
952 | ret = intel_timeline_get_seqno(tl, rq, seqno: &seqno); |
953 | if (ret) |
954 | goto err_free; |
955 | |
956 | dma_fence_init(fence: &rq->fence, ops: &i915_fence_ops, lock: &rq->lock, |
957 | context: tl->fence_context, seqno); |
958 | |
959 | RCU_INIT_POINTER(rq->timeline, tl); |
960 | rq->hwsp_seqno = tl->hwsp_seqno; |
961 | GEM_BUG_ON(__i915_request_is_complete(rq)); |
962 | |
963 | rq->rcustate = get_state_synchronize_rcu(); /* acts as smp_mb() */ |
964 | |
965 | rq->guc_prio = GUC_PRIO_INIT; |
966 | |
967 | /* We bump the ref for the fence chain */ |
968 | i915_sw_fence_reinit(fence: &i915_request_get(rq)->submit); |
969 | i915_sw_fence_reinit(fence: &i915_request_get(rq)->semaphore); |
970 | |
971 | i915_sched_node_reinit(node: &rq->sched); |
972 | |
973 | /* No zalloc, everything must be cleared after use */ |
974 | clear_batch_ptr(rq); |
975 | __rq_init_watchdog(rq); |
976 | assert_capture_list_is_null(rq); |
977 | GEM_BUG_ON(!llist_empty(&rq->execute_cb)); |
978 | GEM_BUG_ON(rq->batch_res); |
979 | |
980 | /* |
981 | * Reserve space in the ring buffer for all the commands required to |
982 | * eventually emit this request. This is to guarantee that the |
983 | * i915_request_add() call can't fail. Note that the reserve may need |
984 | * to be redone if the request is not actually submitted straight |
985 | * away, e.g. because a GPU scheduler has deferred it. |
986 | * |
987 | * Note that due to how we add reserved_space to intel_ring_begin() |
988 | * we need to double our request to ensure that if we need to wrap |
989 | * around inside i915_request_add() there is sufficient space at |
990 | * the beginning of the ring as well. |
991 | */ |
992 | rq->reserved_space = |
993 | 2 * rq->engine->emit_fini_breadcrumb_dw * sizeof(u32); |
994 | |
995 | /* |
996 | * Record the position of the start of the request so that |
997 | * should we detect the updated seqno part-way through the |
998 | * GPU processing the request, we never over-estimate the |
999 | * position of the head. |
1000 | */ |
1001 | rq->head = rq->ring->emit; |
1002 | |
1003 | ret = rq->engine->request_alloc(rq); |
1004 | if (ret) |
1005 | goto err_unwind; |
1006 | |
1007 | rq->infix = rq->ring->emit; /* end of header; start of user payload */ |
1008 | |
1009 | intel_context_mark_active(ce); |
1010 | list_add_tail_rcu(new: &rq->link, head: &tl->requests); |
1011 | |
1012 | return rq; |
1013 | |
1014 | err_unwind: |
1015 | ce->ring->emit = rq->head; |
1016 | |
1017 | /* Make sure we didn't add ourselves to external state before freeing */ |
1018 | GEM_BUG_ON(!list_empty(&rq->sched.signalers_list)); |
1019 | GEM_BUG_ON(!list_empty(&rq->sched.waiters_list)); |
1020 | |
1021 | err_free: |
1022 | kmem_cache_free(s: slab_requests, objp: rq); |
1023 | err_unreserve: |
1024 | intel_context_unpin(ce); |
1025 | return ERR_PTR(error: ret); |
1026 | } |
1027 | |
1028 | struct i915_request * |
1029 | i915_request_create(struct intel_context *ce) |
1030 | { |
1031 | struct i915_request *rq; |
1032 | struct intel_timeline *tl; |
1033 | |
1034 | tl = intel_context_timeline_lock(ce); |
1035 | if (IS_ERR(ptr: tl)) |
1036 | return ERR_CAST(ptr: tl); |
1037 | |
1038 | /* Move our oldest request to the slab-cache (if not in use!) */ |
1039 | rq = list_first_entry(&tl->requests, typeof(*rq), link); |
1040 | if (!list_is_last(list: &rq->link, head: &tl->requests)) |
1041 | i915_request_retire(rq); |
1042 | |
1043 | intel_context_enter(ce); |
1044 | rq = __i915_request_create(ce, GFP_KERNEL); |
1045 | intel_context_exit(ce); /* active reference transferred to request */ |
1046 | if (IS_ERR(ptr: rq)) |
1047 | goto err_unlock; |
1048 | |
1049 | /* Check that we do not interrupt ourselves with a new request */ |
1050 | rq->cookie = lockdep_pin_lock(&tl->mutex); |
1051 | |
1052 | return rq; |
1053 | |
1054 | err_unlock: |
1055 | intel_context_timeline_unlock(tl); |
1056 | return rq; |
1057 | } |
1058 | |
1059 | static int |
1060 | i915_request_await_start(struct i915_request *rq, struct i915_request *signal) |
1061 | { |
1062 | struct dma_fence *fence; |
1063 | int err; |
1064 | |
1065 | if (i915_request_timeline(rq) == rcu_access_pointer(signal->timeline)) |
1066 | return 0; |
1067 | |
1068 | if (i915_request_started(rq: signal)) |
1069 | return 0; |
1070 | |
1071 | /* |
1072 | * The caller holds a reference on @signal, but we do not serialise |
1073 | * against it being retired and removed from the lists. |
1074 | * |
1075 | * We do not hold a reference to the request before @signal, and |
1076 | * so must be very careful to ensure that it is not _recycled_ as |
1077 | * we follow the link backwards. |
1078 | */ |
1079 | fence = NULL; |
1080 | rcu_read_lock(); |
1081 | do { |
1082 | struct list_head *pos = READ_ONCE(signal->link.prev); |
1083 | struct i915_request *prev; |
1084 | |
1085 | /* Confirm signal has not been retired, the link is valid */ |
1086 | if (unlikely(__i915_request_has_started(signal))) |
1087 | break; |
1088 | |
1089 | /* Is signal the earliest request on its timeline? */ |
1090 | if (pos == &rcu_dereference(signal->timeline)->requests) |
1091 | break; |
1092 | |
1093 | /* |
1094 | * Peek at the request before us in the timeline. That |
1095 | * request will only be valid before it is retired, so |
1096 | * after acquiring a reference to it, confirm that it is |
1097 | * still part of the signaler's timeline. |
1098 | */ |
1099 | prev = list_entry(pos, typeof(*prev), link); |
1100 | if (!i915_request_get_rcu(rq: prev)) |
1101 | break; |
1102 | |
1103 | /* After the strong barrier, confirm prev is still attached */ |
1104 | if (unlikely(READ_ONCE(prev->link.next) != &signal->link)) { |
1105 | i915_request_put(rq: prev); |
1106 | break; |
1107 | } |
1108 | |
1109 | fence = &prev->fence; |
1110 | } while (0); |
1111 | rcu_read_unlock(); |
1112 | if (!fence) |
1113 | return 0; |
1114 | |
1115 | err = 0; |
1116 | if (!intel_timeline_sync_is_later(tl: i915_request_timeline(rq), fence)) |
1117 | err = i915_sw_fence_await_dma_fence(fence: &rq->submit, |
1118 | dma: fence, timeout: 0, |
1119 | I915_FENCE_GFP); |
1120 | dma_fence_put(fence); |
1121 | |
1122 | return err; |
1123 | } |
1124 | |
1125 | static intel_engine_mask_t |
1126 | already_busywaiting(struct i915_request *rq) |
1127 | { |
1128 | /* |
1129 | * Polling a semaphore causes bus traffic, delaying other users of |
1130 | * both the GPU and CPU. We want to limit the impact on others, |
1131 | * while taking advantage of early submission to reduce GPU |
1132 | * latency. Therefore we restrict ourselves to not using more |
1133 | * than one semaphore from each source, and not using a semaphore |
1134 | * if we have detected the engine is saturated (i.e. would not be |
1135 | * submitted early and cause bus traffic reading an already passed |
1136 | * semaphore). |
1137 | * |
1138 | * See the are-we-too-late? check in __i915_request_submit(). |
1139 | */ |
1140 | return rq->sched.semaphores | READ_ONCE(rq->engine->saturated); |
1141 | } |
1142 | |
1143 | static int |
1144 | __emit_semaphore_wait(struct i915_request *to, |
1145 | struct i915_request *from, |
1146 | u32 seqno) |
1147 | { |
1148 | const int has_token = GRAPHICS_VER(to->engine->i915) >= 12; |
1149 | u32 hwsp_offset; |
1150 | int len, err; |
1151 | u32 *cs; |
1152 | |
1153 | GEM_BUG_ON(GRAPHICS_VER(to->engine->i915) < 8); |
1154 | GEM_BUG_ON(i915_request_has_initial_breadcrumb(to)); |
1155 | |
1156 | /* We need to pin the signaler's HWSP until we are finished reading. */ |
1157 | err = intel_timeline_read_hwsp(from, until: to, hwsp_offset: &hwsp_offset); |
1158 | if (err) |
1159 | return err; |
1160 | |
1161 | len = 4; |
1162 | if (has_token) |
1163 | len += 2; |
1164 | |
1165 | cs = intel_ring_begin(rq: to, num_dwords: len); |
1166 | if (IS_ERR(ptr: cs)) |
1167 | return PTR_ERR(ptr: cs); |
1168 | |
1169 | /* |
1170 | * Using greater-than-or-equal here means we have to worry |
1171 | * about seqno wraparound. To side step that issue, we swap |
1172 | * the timeline HWSP upon wrapping, so that everyone listening |
1173 | * for the old (pre-wrap) values do not see the much smaller |
1174 | * (post-wrap) values than they were expecting (and so wait |
1175 | * forever). |
1176 | */ |
1177 | *cs++ = (MI_SEMAPHORE_WAIT | |
1178 | MI_SEMAPHORE_GLOBAL_GTT | |
1179 | MI_SEMAPHORE_POLL | |
1180 | MI_SEMAPHORE_SAD_GTE_SDD) + |
1181 | has_token; |
1182 | *cs++ = seqno; |
1183 | *cs++ = hwsp_offset; |
1184 | *cs++ = 0; |
1185 | if (has_token) { |
1186 | *cs++ = 0; |
1187 | *cs++ = MI_NOOP; |
1188 | } |
1189 | |
1190 | intel_ring_advance(rq: to, cs); |
1191 | return 0; |
1192 | } |
1193 | |
1194 | static bool |
1195 | can_use_semaphore_wait(struct i915_request *to, struct i915_request *from) |
1196 | { |
1197 | return to->engine->gt->ggtt == from->engine->gt->ggtt; |
1198 | } |
1199 | |
1200 | static int |
1201 | emit_semaphore_wait(struct i915_request *to, |
1202 | struct i915_request *from, |
1203 | gfp_t gfp) |
1204 | { |
1205 | const intel_engine_mask_t mask = READ_ONCE(from->engine)->mask; |
1206 | struct i915_sw_fence *wait = &to->submit; |
1207 | |
1208 | if (!can_use_semaphore_wait(to, from)) |
1209 | goto await_fence; |
1210 | |
1211 | if (!intel_context_use_semaphores(ce: to->context)) |
1212 | goto await_fence; |
1213 | |
1214 | if (i915_request_has_initial_breadcrumb(rq: to)) |
1215 | goto await_fence; |
1216 | |
1217 | /* |
1218 | * If this or its dependents are waiting on an external fence |
1219 | * that may fail catastrophically, then we want to avoid using |
1220 | * semaphores as they bypass the fence signaling metadata, and we |
1221 | * lose the fence->error propagation. |
1222 | */ |
1223 | if (from->sched.flags & I915_SCHED_HAS_EXTERNAL_CHAIN) |
1224 | goto await_fence; |
1225 | |
1226 | /* Just emit the first semaphore we see as request space is limited. */ |
1227 | if (already_busywaiting(rq: to) & mask) |
1228 | goto await_fence; |
1229 | |
1230 | if (i915_request_await_start(rq: to, signal: from) < 0) |
1231 | goto await_fence; |
1232 | |
1233 | /* Only submit our spinner after the signaler is running! */ |
1234 | if (__await_execution(rq: to, signal: from, gfp)) |
1235 | goto await_fence; |
1236 | |
1237 | if (__emit_semaphore_wait(to, from, seqno: from->fence.seqno)) |
1238 | goto await_fence; |
1239 | |
1240 | to->sched.semaphores |= mask; |
1241 | wait = &to->semaphore; |
1242 | |
1243 | await_fence: |
1244 | return i915_sw_fence_await_dma_fence(fence: wait, |
1245 | dma: &from->fence, timeout: 0, |
1246 | I915_FENCE_GFP); |
1247 | } |
1248 | |
1249 | static bool intel_timeline_sync_has_start(struct intel_timeline *tl, |
1250 | struct dma_fence *fence) |
1251 | { |
1252 | return __intel_timeline_sync_is_later(tl, |
1253 | context: fence->context, |
1254 | seqno: fence->seqno - 1); |
1255 | } |
1256 | |
1257 | static int intel_timeline_sync_set_start(struct intel_timeline *tl, |
1258 | const struct dma_fence *fence) |
1259 | { |
1260 | return __intel_timeline_sync_set(tl, context: fence->context, seqno: fence->seqno - 1); |
1261 | } |
1262 | |
1263 | static int |
1264 | __i915_request_await_execution(struct i915_request *to, |
1265 | struct i915_request *from) |
1266 | { |
1267 | int err; |
1268 | |
1269 | GEM_BUG_ON(intel_context_is_barrier(from->context)); |
1270 | |
1271 | /* Submit both requests at the same time */ |
1272 | err = __await_execution(rq: to, signal: from, I915_FENCE_GFP); |
1273 | if (err) |
1274 | return err; |
1275 | |
1276 | /* Squash repeated depenendices to the same timelines */ |
1277 | if (intel_timeline_sync_has_start(tl: i915_request_timeline(rq: to), |
1278 | fence: &from->fence)) |
1279 | return 0; |
1280 | |
1281 | /* |
1282 | * Wait until the start of this request. |
1283 | * |
1284 | * The execution cb fires when we submit the request to HW. But in |
1285 | * many cases this may be long before the request itself is ready to |
1286 | * run (consider that we submit 2 requests for the same context, where |
1287 | * the request of interest is behind an indefinite spinner). So we hook |
1288 | * up to both to reduce our queues and keep the execution lag minimised |
1289 | * in the worst case, though we hope that the await_start is elided. |
1290 | */ |
1291 | err = i915_request_await_start(rq: to, signal: from); |
1292 | if (err < 0) |
1293 | return err; |
1294 | |
1295 | /* |
1296 | * Ensure both start together [after all semaphores in signal] |
1297 | * |
1298 | * Now that we are queued to the HW at roughly the same time (thanks |
1299 | * to the execute cb) and are ready to run at roughly the same time |
1300 | * (thanks to the await start), our signaler may still be indefinitely |
1301 | * delayed by waiting on a semaphore from a remote engine. If our |
1302 | * signaler depends on a semaphore, so indirectly do we, and we do not |
1303 | * want to start our payload until our signaler also starts theirs. |
1304 | * So we wait. |
1305 | * |
1306 | * However, there is also a second condition for which we need to wait |
1307 | * for the precise start of the signaler. Consider that the signaler |
1308 | * was submitted in a chain of requests following another context |
1309 | * (with just an ordinary intra-engine fence dependency between the |
1310 | * two). In this case the signaler is queued to HW, but not for |
1311 | * immediate execution, and so we must wait until it reaches the |
1312 | * active slot. |
1313 | */ |
1314 | if (can_use_semaphore_wait(to, from) && |
1315 | intel_engine_has_semaphores(engine: to->engine) && |
1316 | !i915_request_has_initial_breadcrumb(rq: to)) { |
1317 | err = __emit_semaphore_wait(to, from, seqno: from->fence.seqno - 1); |
1318 | if (err < 0) |
1319 | return err; |
1320 | } |
1321 | |
1322 | /* Couple the dependency tree for PI on this exposed to->fence */ |
1323 | if (to->engine->sched_engine->schedule) { |
1324 | err = i915_sched_node_add_dependency(node: &to->sched, |
1325 | signal: &from->sched, |
1326 | I915_DEPENDENCY_WEAK); |
1327 | if (err < 0) |
1328 | return err; |
1329 | } |
1330 | |
1331 | return intel_timeline_sync_set_start(tl: i915_request_timeline(rq: to), |
1332 | fence: &from->fence); |
1333 | } |
1334 | |
1335 | static void mark_external(struct i915_request *rq) |
1336 | { |
1337 | /* |
1338 | * The downside of using semaphores is that we lose metadata passing |
1339 | * along the signaling chain. This is particularly nasty when we |
1340 | * need to pass along a fatal error such as EFAULT or EDEADLK. For |
1341 | * fatal errors we want to scrub the request before it is executed, |
1342 | * which means that we cannot preload the request onto HW and have |
1343 | * it wait upon a semaphore. |
1344 | */ |
1345 | rq->sched.flags |= I915_SCHED_HAS_EXTERNAL_CHAIN; |
1346 | } |
1347 | |
1348 | static int |
1349 | __i915_request_await_external(struct i915_request *rq, struct dma_fence *fence) |
1350 | { |
1351 | mark_external(rq); |
1352 | return i915_sw_fence_await_dma_fence(fence: &rq->submit, dma: fence, |
1353 | timeout: i915_fence_context_timeout(i915: rq->i915, |
1354 | context: fence->context), |
1355 | I915_FENCE_GFP); |
1356 | } |
1357 | |
1358 | static int |
1359 | i915_request_await_external(struct i915_request *rq, struct dma_fence *fence) |
1360 | { |
1361 | struct dma_fence *iter; |
1362 | int err = 0; |
1363 | |
1364 | if (!to_dma_fence_chain(fence)) |
1365 | return __i915_request_await_external(rq, fence); |
1366 | |
1367 | dma_fence_chain_for_each(iter, fence) { |
1368 | struct dma_fence_chain *chain = to_dma_fence_chain(fence: iter); |
1369 | |
1370 | if (!dma_fence_is_i915(fence: chain->fence)) { |
1371 | err = __i915_request_await_external(rq, fence: iter); |
1372 | break; |
1373 | } |
1374 | |
1375 | err = i915_request_await_dma_fence(rq, fence: chain->fence); |
1376 | if (err < 0) |
1377 | break; |
1378 | } |
1379 | |
1380 | dma_fence_put(fence: iter); |
1381 | return err; |
1382 | } |
1383 | |
1384 | static inline bool is_parallel_rq(struct i915_request *rq) |
1385 | { |
1386 | return intel_context_is_parallel(ce: rq->context); |
1387 | } |
1388 | |
1389 | static inline struct intel_context *request_to_parent(struct i915_request *rq) |
1390 | { |
1391 | return intel_context_to_parent(ce: rq->context); |
1392 | } |
1393 | |
1394 | static bool is_same_parallel_context(struct i915_request *to, |
1395 | struct i915_request *from) |
1396 | { |
1397 | if (is_parallel_rq(rq: to)) |
1398 | return request_to_parent(rq: to) == request_to_parent(rq: from); |
1399 | |
1400 | return false; |
1401 | } |
1402 | |
1403 | int |
1404 | i915_request_await_execution(struct i915_request *rq, |
1405 | struct dma_fence *fence) |
1406 | { |
1407 | struct dma_fence **child = &fence; |
1408 | unsigned int nchild = 1; |
1409 | int ret; |
1410 | |
1411 | if (dma_fence_is_array(fence)) { |
1412 | struct dma_fence_array *array = to_dma_fence_array(fence); |
1413 | |
1414 | /* XXX Error for signal-on-any fence arrays */ |
1415 | |
1416 | child = array->fences; |
1417 | nchild = array->num_fences; |
1418 | GEM_BUG_ON(!nchild); |
1419 | } |
1420 | |
1421 | do { |
1422 | fence = *child++; |
1423 | if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) |
1424 | continue; |
1425 | |
1426 | if (fence->context == rq->fence.context) |
1427 | continue; |
1428 | |
1429 | /* |
1430 | * We don't squash repeated fence dependencies here as we |
1431 | * want to run our callback in all cases. |
1432 | */ |
1433 | |
1434 | if (dma_fence_is_i915(fence)) { |
1435 | if (is_same_parallel_context(to: rq, from: to_request(fence))) |
1436 | continue; |
1437 | ret = __i915_request_await_execution(to: rq, |
1438 | from: to_request(fence)); |
1439 | } else { |
1440 | ret = i915_request_await_external(rq, fence); |
1441 | } |
1442 | if (ret < 0) |
1443 | return ret; |
1444 | } while (--nchild); |
1445 | |
1446 | return 0; |
1447 | } |
1448 | |
1449 | static int |
1450 | await_request_submit(struct i915_request *to, struct i915_request *from) |
1451 | { |
1452 | /* |
1453 | * If we are waiting on a virtual engine, then it may be |
1454 | * constrained to execute on a single engine *prior* to submission. |
1455 | * When it is submitted, it will be first submitted to the virtual |
1456 | * engine and then passed to the physical engine. We cannot allow |
1457 | * the waiter to be submitted immediately to the physical engine |
1458 | * as it may then bypass the virtual request. |
1459 | */ |
1460 | if (to->engine == READ_ONCE(from->engine)) |
1461 | return i915_sw_fence_await_sw_fence_gfp(fence: &to->submit, |
1462 | after: &from->submit, |
1463 | I915_FENCE_GFP); |
1464 | else |
1465 | return __i915_request_await_execution(to, from); |
1466 | } |
1467 | |
1468 | static int |
1469 | i915_request_await_request(struct i915_request *to, struct i915_request *from) |
1470 | { |
1471 | int ret; |
1472 | |
1473 | GEM_BUG_ON(to == from); |
1474 | GEM_BUG_ON(to->timeline == from->timeline); |
1475 | |
1476 | if (i915_request_completed(rq: from)) { |
1477 | i915_sw_fence_set_error_once(fence: &to->submit, error: from->fence.error); |
1478 | return 0; |
1479 | } |
1480 | |
1481 | if (to->engine->sched_engine->schedule) { |
1482 | ret = i915_sched_node_add_dependency(node: &to->sched, |
1483 | signal: &from->sched, |
1484 | I915_DEPENDENCY_EXTERNAL); |
1485 | if (ret < 0) |
1486 | return ret; |
1487 | } |
1488 | |
1489 | if (!intel_engine_uses_guc(engine: to->engine) && |
1490 | is_power_of_2(n: to->execution_mask | READ_ONCE(from->execution_mask))) |
1491 | ret = await_request_submit(to, from); |
1492 | else |
1493 | ret = emit_semaphore_wait(to, from, I915_FENCE_GFP); |
1494 | if (ret < 0) |
1495 | return ret; |
1496 | |
1497 | return 0; |
1498 | } |
1499 | |
1500 | int |
1501 | i915_request_await_dma_fence(struct i915_request *rq, struct dma_fence *fence) |
1502 | { |
1503 | struct dma_fence **child = &fence; |
1504 | unsigned int nchild = 1; |
1505 | int ret; |
1506 | |
1507 | /* |
1508 | * Note that if the fence-array was created in signal-on-any mode, |
1509 | * we should *not* decompose it into its individual fences. However, |
1510 | * we don't currently store which mode the fence-array is operating |
1511 | * in. Fortunately, the only user of signal-on-any is private to |
1512 | * amdgpu and we should not see any incoming fence-array from |
1513 | * sync-file being in signal-on-any mode. |
1514 | */ |
1515 | if (dma_fence_is_array(fence)) { |
1516 | struct dma_fence_array *array = to_dma_fence_array(fence); |
1517 | |
1518 | child = array->fences; |
1519 | nchild = array->num_fences; |
1520 | GEM_BUG_ON(!nchild); |
1521 | } |
1522 | |
1523 | do { |
1524 | fence = *child++; |
1525 | if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) |
1526 | continue; |
1527 | |
1528 | /* |
1529 | * Requests on the same timeline are explicitly ordered, along |
1530 | * with their dependencies, by i915_request_add() which ensures |
1531 | * that requests are submitted in-order through each ring. |
1532 | */ |
1533 | if (fence->context == rq->fence.context) |
1534 | continue; |
1535 | |
1536 | /* Squash repeated waits to the same timelines */ |
1537 | if (fence->context && |
1538 | intel_timeline_sync_is_later(tl: i915_request_timeline(rq), |
1539 | fence)) |
1540 | continue; |
1541 | |
1542 | if (dma_fence_is_i915(fence)) { |
1543 | if (is_same_parallel_context(to: rq, from: to_request(fence))) |
1544 | continue; |
1545 | ret = i915_request_await_request(to: rq, from: to_request(fence)); |
1546 | } else { |
1547 | ret = i915_request_await_external(rq, fence); |
1548 | } |
1549 | if (ret < 0) |
1550 | return ret; |
1551 | |
1552 | /* Record the latest fence used against each timeline */ |
1553 | if (fence->context) |
1554 | intel_timeline_sync_set(tl: i915_request_timeline(rq), |
1555 | fence); |
1556 | } while (--nchild); |
1557 | |
1558 | return 0; |
1559 | } |
1560 | |
1561 | /** |
1562 | * i915_request_await_deps - set this request to (async) wait upon a struct |
1563 | * i915_deps dma_fence collection |
1564 | * @rq: request we are wishing to use |
1565 | * @deps: The struct i915_deps containing the dependencies. |
1566 | * |
1567 | * Returns 0 if successful, negative error code on error. |
1568 | */ |
1569 | int i915_request_await_deps(struct i915_request *rq, const struct i915_deps *deps) |
1570 | { |
1571 | int i, err; |
1572 | |
1573 | for (i = 0; i < deps->num_deps; ++i) { |
1574 | err = i915_request_await_dma_fence(rq, fence: deps->fences[i]); |
1575 | if (err) |
1576 | return err; |
1577 | } |
1578 | |
1579 | return 0; |
1580 | } |
1581 | |
1582 | /** |
1583 | * i915_request_await_object - set this request to (async) wait upon a bo |
1584 | * @to: request we are wishing to use |
1585 | * @obj: object which may be in use on another ring. |
1586 | * @write: whether the wait is on behalf of a writer |
1587 | * |
1588 | * This code is meant to abstract object synchronization with the GPU. |
1589 | * Conceptually we serialise writes between engines inside the GPU. |
1590 | * We only allow one engine to write into a buffer at any time, but |
1591 | * multiple readers. To ensure each has a coherent view of memory, we must: |
1592 | * |
1593 | * - If there is an outstanding write request to the object, the new |
1594 | * request must wait for it to complete (either CPU or in hw, requests |
1595 | * on the same ring will be naturally ordered). |
1596 | * |
1597 | * - If we are a write request (pending_write_domain is set), the new |
1598 | * request must wait for outstanding read requests to complete. |
1599 | * |
1600 | * Returns 0 if successful, else propagates up the lower layer error. |
1601 | */ |
1602 | int |
1603 | i915_request_await_object(struct i915_request *to, |
1604 | struct drm_i915_gem_object *obj, |
1605 | bool write) |
1606 | { |
1607 | struct dma_resv_iter cursor; |
1608 | struct dma_fence *fence; |
1609 | int ret = 0; |
1610 | |
1611 | dma_resv_for_each_fence(&cursor, obj->base.resv, |
1612 | dma_resv_usage_rw(write), fence) { |
1613 | ret = i915_request_await_dma_fence(rq: to, fence); |
1614 | if (ret) |
1615 | break; |
1616 | } |
1617 | |
1618 | return ret; |
1619 | } |
1620 | |
1621 | static void i915_request_await_huc(struct i915_request *rq) |
1622 | { |
1623 | struct intel_huc *huc = &rq->context->engine->gt->uc.huc; |
1624 | |
1625 | /* don't stall kernel submissions! */ |
1626 | if (!rcu_access_pointer(rq->context->gem_context)) |
1627 | return; |
1628 | |
1629 | if (intel_huc_wait_required(huc)) |
1630 | i915_sw_fence_await_sw_fence(fence: &rq->submit, |
1631 | after: &huc->delayed_load.fence, |
1632 | wq: &rq->hucq); |
1633 | } |
1634 | |
1635 | static struct i915_request * |
1636 | __i915_request_ensure_parallel_ordering(struct i915_request *rq, |
1637 | struct intel_timeline *timeline) |
1638 | { |
1639 | struct i915_request *prev; |
1640 | |
1641 | GEM_BUG_ON(!is_parallel_rq(rq)); |
1642 | |
1643 | prev = request_to_parent(rq)->parallel.last_rq; |
1644 | if (prev) { |
1645 | if (!__i915_request_is_complete(rq: prev)) { |
1646 | i915_sw_fence_await_sw_fence(fence: &rq->submit, |
1647 | after: &prev->submit, |
1648 | wq: &rq->submitq); |
1649 | |
1650 | if (rq->engine->sched_engine->schedule) |
1651 | __i915_sched_node_add_dependency(node: &rq->sched, |
1652 | signal: &prev->sched, |
1653 | dep: &rq->dep, |
1654 | flags: 0); |
1655 | } |
1656 | i915_request_put(rq: prev); |
1657 | } |
1658 | |
1659 | request_to_parent(rq)->parallel.last_rq = i915_request_get(rq); |
1660 | |
1661 | /* |
1662 | * Users have to put a reference potentially got by |
1663 | * __i915_active_fence_set() to the returned request |
1664 | * when no longer needed |
1665 | */ |
1666 | return to_request(fence: __i915_active_fence_set(active: &timeline->last_request, |
1667 | fence: &rq->fence)); |
1668 | } |
1669 | |
1670 | static struct i915_request * |
1671 | __i915_request_ensure_ordering(struct i915_request *rq, |
1672 | struct intel_timeline *timeline) |
1673 | { |
1674 | struct i915_request *prev; |
1675 | |
1676 | GEM_BUG_ON(is_parallel_rq(rq)); |
1677 | |
1678 | prev = to_request(fence: __i915_active_fence_set(active: &timeline->last_request, |
1679 | fence: &rq->fence)); |
1680 | |
1681 | if (prev && !__i915_request_is_complete(rq: prev)) { |
1682 | bool uses_guc = intel_engine_uses_guc(engine: rq->engine); |
1683 | bool pow2 = is_power_of_2(READ_ONCE(prev->engine)->mask | |
1684 | rq->engine->mask); |
1685 | bool same_context = prev->context == rq->context; |
1686 | |
1687 | /* |
1688 | * The requests are supposed to be kept in order. However, |
1689 | * we need to be wary in case the timeline->last_request |
1690 | * is used as a barrier for external modification to this |
1691 | * context. |
1692 | */ |
1693 | GEM_BUG_ON(same_context && |
1694 | i915_seqno_passed(prev->fence.seqno, |
1695 | rq->fence.seqno)); |
1696 | |
1697 | if ((same_context && uses_guc) || (!uses_guc && pow2)) |
1698 | i915_sw_fence_await_sw_fence(fence: &rq->submit, |
1699 | after: &prev->submit, |
1700 | wq: &rq->submitq); |
1701 | else |
1702 | __i915_sw_fence_await_dma_fence(fence: &rq->submit, |
1703 | dma: &prev->fence, |
1704 | cb: &rq->dmaq); |
1705 | if (rq->engine->sched_engine->schedule) |
1706 | __i915_sched_node_add_dependency(node: &rq->sched, |
1707 | signal: &prev->sched, |
1708 | dep: &rq->dep, |
1709 | flags: 0); |
1710 | } |
1711 | |
1712 | /* |
1713 | * Users have to put the reference to prev potentially got |
1714 | * by __i915_active_fence_set() when no longer needed |
1715 | */ |
1716 | return prev; |
1717 | } |
1718 | |
1719 | static struct i915_request * |
1720 | __i915_request_add_to_timeline(struct i915_request *rq) |
1721 | { |
1722 | struct intel_timeline *timeline = i915_request_timeline(rq); |
1723 | struct i915_request *prev; |
1724 | |
1725 | /* |
1726 | * Media workloads may require HuC, so stall them until HuC loading is |
1727 | * complete. Note that HuC not being loaded when a user submission |
1728 | * arrives can only happen when HuC is loaded via GSC and in that case |
1729 | * we still expect the window between us starting to accept submissions |
1730 | * and HuC loading completion to be small (a few hundred ms). |
1731 | */ |
1732 | if (rq->engine->class == VIDEO_DECODE_CLASS) |
1733 | i915_request_await_huc(rq); |
1734 | |
1735 | /* |
1736 | * Dependency tracking and request ordering along the timeline |
1737 | * is special cased so that we can eliminate redundant ordering |
1738 | * operations while building the request (we know that the timeline |
1739 | * itself is ordered, and here we guarantee it). |
1740 | * |
1741 | * As we know we will need to emit tracking along the timeline, |
1742 | * we embed the hooks into our request struct -- at the cost of |
1743 | * having to have specialised no-allocation interfaces (which will |
1744 | * be beneficial elsewhere). |
1745 | * |
1746 | * A second benefit to open-coding i915_request_await_request is |
1747 | * that we can apply a slight variant of the rules specialised |
1748 | * for timelines that jump between engines (such as virtual engines). |
1749 | * If we consider the case of virtual engine, we must emit a dma-fence |
1750 | * to prevent scheduling of the second request until the first is |
1751 | * complete (to maximise our greedy late load balancing) and this |
1752 | * precludes optimising to use semaphores serialisation of a single |
1753 | * timeline across engines. |
1754 | * |
1755 | * We do not order parallel submission requests on the timeline as each |
1756 | * parallel submission context has its own timeline and the ordering |
1757 | * rules for parallel requests are that they must be submitted in the |
1758 | * order received from the execbuf IOCTL. So rather than using the |
1759 | * timeline we store a pointer to last request submitted in the |
1760 | * relationship in the gem context and insert a submission fence |
1761 | * between that request and request passed into this function or |
1762 | * alternatively we use completion fence if gem context has a single |
1763 | * timeline and this is the first submission of an execbuf IOCTL. |
1764 | */ |
1765 | if (likely(!is_parallel_rq(rq))) |
1766 | prev = __i915_request_ensure_ordering(rq, timeline); |
1767 | else |
1768 | prev = __i915_request_ensure_parallel_ordering(rq, timeline); |
1769 | if (prev) |
1770 | i915_request_put(rq: prev); |
1771 | |
1772 | /* |
1773 | * Make sure that no request gazumped us - if it was allocated after |
1774 | * our i915_request_alloc() and called __i915_request_add() before |
1775 | * us, the timeline will hold its seqno which is later than ours. |
1776 | */ |
1777 | GEM_BUG_ON(timeline->seqno != rq->fence.seqno); |
1778 | |
1779 | return prev; |
1780 | } |
1781 | |
1782 | /* |
1783 | * NB: This function is not allowed to fail. Doing so would mean the the |
1784 | * request is not being tracked for completion but the work itself is |
1785 | * going to happen on the hardware. This would be a Bad Thing(tm). |
1786 | */ |
1787 | struct i915_request *__i915_request_commit(struct i915_request *rq) |
1788 | { |
1789 | struct intel_engine_cs *engine = rq->engine; |
1790 | struct intel_ring *ring = rq->ring; |
1791 | u32 *cs; |
1792 | |
1793 | RQ_TRACE(rq, "\n" ); |
1794 | |
1795 | /* |
1796 | * To ensure that this call will not fail, space for its emissions |
1797 | * should already have been reserved in the ring buffer. Let the ring |
1798 | * know that it is time to use that space up. |
1799 | */ |
1800 | GEM_BUG_ON(rq->reserved_space > ring->space); |
1801 | rq->reserved_space = 0; |
1802 | rq->emitted_jiffies = jiffies; |
1803 | |
1804 | /* |
1805 | * Record the position of the start of the breadcrumb so that |
1806 | * should we detect the updated seqno part-way through the |
1807 | * GPU processing the request, we never over-estimate the |
1808 | * position of the ring's HEAD. |
1809 | */ |
1810 | cs = intel_ring_begin(rq, num_dwords: engine->emit_fini_breadcrumb_dw); |
1811 | GEM_BUG_ON(IS_ERR(cs)); |
1812 | rq->postfix = intel_ring_offset(rq, addr: cs); |
1813 | |
1814 | return __i915_request_add_to_timeline(rq); |
1815 | } |
1816 | |
1817 | void __i915_request_queue_bh(struct i915_request *rq) |
1818 | { |
1819 | i915_sw_fence_commit(fence: &rq->semaphore); |
1820 | i915_sw_fence_commit(fence: &rq->submit); |
1821 | } |
1822 | |
1823 | void __i915_request_queue(struct i915_request *rq, |
1824 | const struct i915_sched_attr *attr) |
1825 | { |
1826 | /* |
1827 | * Let the backend know a new request has arrived that may need |
1828 | * to adjust the existing execution schedule due to a high priority |
1829 | * request - i.e. we may want to preempt the current request in order |
1830 | * to run a high priority dependency chain *before* we can execute this |
1831 | * request. |
1832 | * |
1833 | * This is called before the request is ready to run so that we can |
1834 | * decide whether to preempt the entire chain so that it is ready to |
1835 | * run at the earliest possible convenience. |
1836 | */ |
1837 | if (attr && rq->engine->sched_engine->schedule) |
1838 | rq->engine->sched_engine->schedule(rq, attr); |
1839 | |
1840 | local_bh_disable(); |
1841 | __i915_request_queue_bh(rq); |
1842 | local_bh_enable(); /* kick tasklets */ |
1843 | } |
1844 | |
1845 | void i915_request_add(struct i915_request *rq) |
1846 | { |
1847 | struct intel_timeline * const tl = i915_request_timeline(rq); |
1848 | struct i915_sched_attr attr = {}; |
1849 | struct i915_gem_context *ctx; |
1850 | |
1851 | lockdep_assert_held(&tl->mutex); |
1852 | lockdep_unpin_lock(&tl->mutex, rq->cookie); |
1853 | |
1854 | trace_i915_request_add(rq); |
1855 | __i915_request_commit(rq); |
1856 | |
1857 | /* XXX placeholder for selftests */ |
1858 | rcu_read_lock(); |
1859 | ctx = rcu_dereference(rq->context->gem_context); |
1860 | if (ctx) |
1861 | attr = ctx->sched; |
1862 | rcu_read_unlock(); |
1863 | |
1864 | __i915_request_queue(rq, attr: &attr); |
1865 | |
1866 | mutex_unlock(lock: &tl->mutex); |
1867 | } |
1868 | |
1869 | static unsigned long local_clock_ns(unsigned int *cpu) |
1870 | { |
1871 | unsigned long t; |
1872 | |
1873 | /* |
1874 | * Cheaply and approximately convert from nanoseconds to microseconds. |
1875 | * The result and subsequent calculations are also defined in the same |
1876 | * approximate microseconds units. The principal source of timing |
1877 | * error here is from the simple truncation. |
1878 | * |
1879 | * Note that local_clock() is only defined wrt to the current CPU; |
1880 | * the comparisons are no longer valid if we switch CPUs. Instead of |
1881 | * blocking preemption for the entire busywait, we can detect the CPU |
1882 | * switch and use that as indicator of system load and a reason to |
1883 | * stop busywaiting, see busywait_stop(). |
1884 | */ |
1885 | *cpu = get_cpu(); |
1886 | t = local_clock(); |
1887 | put_cpu(); |
1888 | |
1889 | return t; |
1890 | } |
1891 | |
1892 | static bool busywait_stop(unsigned long timeout, unsigned int cpu) |
1893 | { |
1894 | unsigned int this_cpu; |
1895 | |
1896 | if (time_after(local_clock_ns(&this_cpu), timeout)) |
1897 | return true; |
1898 | |
1899 | return this_cpu != cpu; |
1900 | } |
1901 | |
1902 | static bool __i915_spin_request(struct i915_request * const rq, int state) |
1903 | { |
1904 | unsigned long timeout_ns; |
1905 | unsigned int cpu; |
1906 | |
1907 | /* |
1908 | * Only wait for the request if we know it is likely to complete. |
1909 | * |
1910 | * We don't track the timestamps around requests, nor the average |
1911 | * request length, so we do not have a good indicator that this |
1912 | * request will complete within the timeout. What we do know is the |
1913 | * order in which requests are executed by the context and so we can |
1914 | * tell if the request has been started. If the request is not even |
1915 | * running yet, it is a fair assumption that it will not complete |
1916 | * within our relatively short timeout. |
1917 | */ |
1918 | if (!i915_request_is_running(rq)) |
1919 | return false; |
1920 | |
1921 | /* |
1922 | * When waiting for high frequency requests, e.g. during synchronous |
1923 | * rendering split between the CPU and GPU, the finite amount of time |
1924 | * required to set up the irq and wait upon it limits the response |
1925 | * rate. By busywaiting on the request completion for a short while we |
1926 | * can service the high frequency waits as quick as possible. However, |
1927 | * if it is a slow request, we want to sleep as quickly as possible. |
1928 | * The tradeoff between waiting and sleeping is roughly the time it |
1929 | * takes to sleep on a request, on the order of a microsecond. |
1930 | */ |
1931 | |
1932 | timeout_ns = READ_ONCE(rq->engine->props.max_busywait_duration_ns); |
1933 | timeout_ns += local_clock_ns(cpu: &cpu); |
1934 | do { |
1935 | if (dma_fence_is_signaled(fence: &rq->fence)) |
1936 | return true; |
1937 | |
1938 | if (signal_pending_state(state, current)) |
1939 | break; |
1940 | |
1941 | if (busywait_stop(timeout: timeout_ns, cpu)) |
1942 | break; |
1943 | |
1944 | cpu_relax(); |
1945 | } while (!need_resched()); |
1946 | |
1947 | return false; |
1948 | } |
1949 | |
1950 | struct request_wait { |
1951 | struct dma_fence_cb cb; |
1952 | struct task_struct *tsk; |
1953 | }; |
1954 | |
1955 | static void request_wait_wake(struct dma_fence *fence, struct dma_fence_cb *cb) |
1956 | { |
1957 | struct request_wait *wait = container_of(cb, typeof(*wait), cb); |
1958 | |
1959 | wake_up_process(fetch_and_zero(&wait->tsk)); |
1960 | } |
1961 | |
1962 | /** |
1963 | * i915_request_wait_timeout - wait until execution of request has finished |
1964 | * @rq: the request to wait upon |
1965 | * @flags: how to wait |
1966 | * @timeout: how long to wait in jiffies |
1967 | * |
1968 | * i915_request_wait_timeout() waits for the request to be completed, for a |
1969 | * maximum of @timeout jiffies (with MAX_SCHEDULE_TIMEOUT implying an |
1970 | * unbounded wait). |
1971 | * |
1972 | * Returns the remaining time (in jiffies) if the request completed, which may |
1973 | * be zero if the request is unfinished after the timeout expires. |
1974 | * If the timeout is 0, it will return 1 if the fence is signaled. |
1975 | * |
1976 | * May return -EINTR is called with I915_WAIT_INTERRUPTIBLE and a signal is |
1977 | * pending before the request completes. |
1978 | * |
1979 | * NOTE: This function has the same wait semantics as dma-fence. |
1980 | */ |
1981 | long i915_request_wait_timeout(struct i915_request *rq, |
1982 | unsigned int flags, |
1983 | long timeout) |
1984 | { |
1985 | const int state = flags & I915_WAIT_INTERRUPTIBLE ? |
1986 | TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE; |
1987 | struct request_wait wait; |
1988 | |
1989 | might_sleep(); |
1990 | GEM_BUG_ON(timeout < 0); |
1991 | |
1992 | if (dma_fence_is_signaled(fence: &rq->fence)) |
1993 | return timeout ?: 1; |
1994 | |
1995 | if (!timeout) |
1996 | return -ETIME; |
1997 | |
1998 | trace_i915_request_wait_begin(rq, flags); |
1999 | |
2000 | /* |
2001 | * We must never wait on the GPU while holding a lock as we |
2002 | * may need to perform a GPU reset. So while we don't need to |
2003 | * serialise wait/reset with an explicit lock, we do want |
2004 | * lockdep to detect potential dependency cycles. |
2005 | */ |
2006 | mutex_acquire(&rq->engine->gt->reset.mutex.dep_map, 0, 0, _THIS_IP_); |
2007 | |
2008 | /* |
2009 | * Optimistic spin before touching IRQs. |
2010 | * |
2011 | * We may use a rather large value here to offset the penalty of |
2012 | * switching away from the active task. Frequently, the client will |
2013 | * wait upon an old swapbuffer to throttle itself to remain within a |
2014 | * frame of the gpu. If the client is running in lockstep with the gpu, |
2015 | * then it should not be waiting long at all, and a sleep now will incur |
2016 | * extra scheduler latency in producing the next frame. To try to |
2017 | * avoid adding the cost of enabling/disabling the interrupt to the |
2018 | * short wait, we first spin to see if the request would have completed |
2019 | * in the time taken to setup the interrupt. |
2020 | * |
2021 | * We need upto 5us to enable the irq, and upto 20us to hide the |
2022 | * scheduler latency of a context switch, ignoring the secondary |
2023 | * impacts from a context switch such as cache eviction. |
2024 | * |
2025 | * The scheme used for low-latency IO is called "hybrid interrupt |
2026 | * polling". The suggestion there is to sleep until just before you |
2027 | * expect to be woken by the device interrupt and then poll for its |
2028 | * completion. That requires having a good predictor for the request |
2029 | * duration, which we currently lack. |
2030 | */ |
2031 | if (CONFIG_DRM_I915_MAX_REQUEST_BUSYWAIT && |
2032 | __i915_spin_request(rq, state)) |
2033 | goto out; |
2034 | |
2035 | /* |
2036 | * This client is about to stall waiting for the GPU. In many cases |
2037 | * this is undesirable and limits the throughput of the system, as |
2038 | * many clients cannot continue processing user input/output whilst |
2039 | * blocked. RPS autotuning may take tens of milliseconds to respond |
2040 | * to the GPU load and thus incurs additional latency for the client. |
2041 | * We can circumvent that by promoting the GPU frequency to maximum |
2042 | * before we sleep. This makes the GPU throttle up much more quickly |
2043 | * (good for benchmarks and user experience, e.g. window animations), |
2044 | * but at a cost of spending more power processing the workload |
2045 | * (bad for battery). |
2046 | */ |
2047 | if (flags & I915_WAIT_PRIORITY && !i915_request_started(rq)) |
2048 | intel_rps_boost(rq); |
2049 | |
2050 | wait.tsk = current; |
2051 | if (dma_fence_add_callback(fence: &rq->fence, cb: &wait.cb, func: request_wait_wake)) |
2052 | goto out; |
2053 | |
2054 | /* |
2055 | * Flush the submission tasklet, but only if it may help this request. |
2056 | * |
2057 | * We sometimes experience some latency between the HW interrupts and |
2058 | * tasklet execution (mostly due to ksoftirqd latency, but it can also |
2059 | * be due to lazy CS events), so lets run the tasklet manually if there |
2060 | * is a chance it may submit this request. If the request is not ready |
2061 | * to run, as it is waiting for other fences to be signaled, flushing |
2062 | * the tasklet is busy work without any advantage for this client. |
2063 | * |
2064 | * If the HW is being lazy, this is the last chance before we go to |
2065 | * sleep to catch any pending events. We will check periodically in |
2066 | * the heartbeat to flush the submission tasklets as a last resort |
2067 | * for unhappy HW. |
2068 | */ |
2069 | if (i915_request_is_ready(rq)) |
2070 | __intel_engine_flush_submission(engine: rq->engine, sync: false); |
2071 | |
2072 | for (;;) { |
2073 | set_current_state(state); |
2074 | |
2075 | if (dma_fence_is_signaled(fence: &rq->fence)) |
2076 | break; |
2077 | |
2078 | if (signal_pending_state(state, current)) { |
2079 | timeout = -ERESTARTSYS; |
2080 | break; |
2081 | } |
2082 | |
2083 | if (!timeout) { |
2084 | timeout = -ETIME; |
2085 | break; |
2086 | } |
2087 | |
2088 | timeout = io_schedule_timeout(timeout); |
2089 | } |
2090 | __set_current_state(TASK_RUNNING); |
2091 | |
2092 | if (READ_ONCE(wait.tsk)) |
2093 | dma_fence_remove_callback(fence: &rq->fence, cb: &wait.cb); |
2094 | GEM_BUG_ON(!list_empty(&wait.cb.node)); |
2095 | |
2096 | out: |
2097 | mutex_release(&rq->engine->gt->reset.mutex.dep_map, _THIS_IP_); |
2098 | trace_i915_request_wait_end(rq); |
2099 | return timeout; |
2100 | } |
2101 | |
2102 | /** |
2103 | * i915_request_wait - wait until execution of request has finished |
2104 | * @rq: the request to wait upon |
2105 | * @flags: how to wait |
2106 | * @timeout: how long to wait in jiffies |
2107 | * |
2108 | * i915_request_wait() waits for the request to be completed, for a |
2109 | * maximum of @timeout jiffies (with MAX_SCHEDULE_TIMEOUT implying an |
2110 | * unbounded wait). |
2111 | * |
2112 | * Returns the remaining time (in jiffies) if the request completed, which may |
2113 | * be zero or -ETIME if the request is unfinished after the timeout expires. |
2114 | * May return -EINTR is called with I915_WAIT_INTERRUPTIBLE and a signal is |
2115 | * pending before the request completes. |
2116 | * |
2117 | * NOTE: This function behaves differently from dma-fence wait semantics for |
2118 | * timeout = 0. It returns 0 on success, and -ETIME if not signaled. |
2119 | */ |
2120 | long i915_request_wait(struct i915_request *rq, |
2121 | unsigned int flags, |
2122 | long timeout) |
2123 | { |
2124 | long ret = i915_request_wait_timeout(rq, flags, timeout); |
2125 | |
2126 | if (!ret) |
2127 | return -ETIME; |
2128 | |
2129 | if (ret > 0 && !timeout) |
2130 | return 0; |
2131 | |
2132 | return ret; |
2133 | } |
2134 | |
2135 | static int print_sched_attr(const struct i915_sched_attr *attr, |
2136 | char *buf, int x, int len) |
2137 | { |
2138 | if (attr->priority == I915_PRIORITY_INVALID) |
2139 | return x; |
2140 | |
2141 | x += snprintf(buf: buf + x, size: len - x, |
2142 | fmt: " prio=%d" , attr->priority); |
2143 | |
2144 | return x; |
2145 | } |
2146 | |
2147 | static char queue_status(const struct i915_request *rq) |
2148 | { |
2149 | if (i915_request_is_active(rq)) |
2150 | return 'E'; |
2151 | |
2152 | if (i915_request_is_ready(rq)) |
2153 | return intel_engine_is_virtual(engine: rq->engine) ? 'V' : 'R'; |
2154 | |
2155 | return 'U'; |
2156 | } |
2157 | |
2158 | static const char *run_status(const struct i915_request *rq) |
2159 | { |
2160 | if (__i915_request_is_complete(rq)) |
2161 | return "!" ; |
2162 | |
2163 | if (__i915_request_has_started(rq)) |
2164 | return "*" ; |
2165 | |
2166 | if (!i915_sw_fence_signaled(fence: &rq->semaphore)) |
2167 | return "&" ; |
2168 | |
2169 | return "" ; |
2170 | } |
2171 | |
2172 | static const char *fence_status(const struct i915_request *rq) |
2173 | { |
2174 | if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &rq->fence.flags)) |
2175 | return "+" ; |
2176 | |
2177 | if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &rq->fence.flags)) |
2178 | return "-" ; |
2179 | |
2180 | return "" ; |
2181 | } |
2182 | |
2183 | void i915_request_show(struct drm_printer *m, |
2184 | const struct i915_request *rq, |
2185 | const char *prefix, |
2186 | int indent) |
2187 | { |
2188 | const char *name = rq->fence.ops->get_timeline_name((struct dma_fence *)&rq->fence); |
2189 | char buf[80] = "" ; |
2190 | int x = 0; |
2191 | |
2192 | /* |
2193 | * The prefix is used to show the queue status, for which we use |
2194 | * the following flags: |
2195 | * |
2196 | * U [Unready] |
2197 | * - initial status upon being submitted by the user |
2198 | * |
2199 | * - the request is not ready for execution as it is waiting |
2200 | * for external fences |
2201 | * |
2202 | * R [Ready] |
2203 | * - all fences the request was waiting on have been signaled, |
2204 | * and the request is now ready for execution and will be |
2205 | * in a backend queue |
2206 | * |
2207 | * - a ready request may still need to wait on semaphores |
2208 | * [internal fences] |
2209 | * |
2210 | * V [Ready/virtual] |
2211 | * - same as ready, but queued over multiple backends |
2212 | * |
2213 | * E [Executing] |
2214 | * - the request has been transferred from the backend queue and |
2215 | * submitted for execution on HW |
2216 | * |
2217 | * - a completed request may still be regarded as executing, its |
2218 | * status may not be updated until it is retired and removed |
2219 | * from the lists |
2220 | */ |
2221 | |
2222 | x = print_sched_attr(attr: &rq->sched.attr, buf, x, len: sizeof(buf)); |
2223 | |
2224 | drm_printf(p: m, f: "%s%.*s%c %llx:%lld%s%s %s @ %dms: %s\n" , |
2225 | prefix, indent, " " , |
2226 | queue_status(rq), |
2227 | rq->fence.context, rq->fence.seqno, |
2228 | run_status(rq), |
2229 | fence_status(rq), |
2230 | buf, |
2231 | jiffies_to_msecs(j: jiffies - rq->emitted_jiffies), |
2232 | name); |
2233 | } |
2234 | |
2235 | static bool engine_match_ring(struct intel_engine_cs *engine, struct i915_request *rq) |
2236 | { |
2237 | u32 ring = ENGINE_READ(engine, RING_START); |
2238 | |
2239 | return ring == i915_ggtt_offset(vma: rq->ring->vma); |
2240 | } |
2241 | |
2242 | static bool match_ring(struct i915_request *rq) |
2243 | { |
2244 | struct intel_engine_cs *engine; |
2245 | bool found; |
2246 | int i; |
2247 | |
2248 | if (!intel_engine_is_virtual(engine: rq->engine)) |
2249 | return engine_match_ring(engine: rq->engine, rq); |
2250 | |
2251 | found = false; |
2252 | i = 0; |
2253 | while ((engine = intel_engine_get_sibling(engine: rq->engine, sibling: i++))) { |
2254 | found = engine_match_ring(engine, rq); |
2255 | if (found) |
2256 | break; |
2257 | } |
2258 | |
2259 | return found; |
2260 | } |
2261 | |
2262 | enum i915_request_state i915_test_request_state(struct i915_request *rq) |
2263 | { |
2264 | if (i915_request_completed(rq)) |
2265 | return I915_REQUEST_COMPLETE; |
2266 | |
2267 | if (!i915_request_started(rq)) |
2268 | return I915_REQUEST_PENDING; |
2269 | |
2270 | if (match_ring(rq)) |
2271 | return I915_REQUEST_ACTIVE; |
2272 | |
2273 | return I915_REQUEST_QUEUED; |
2274 | } |
2275 | |
2276 | #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) |
2277 | #include "selftests/mock_request.c" |
2278 | #include "selftests/i915_request.c" |
2279 | #endif |
2280 | |
2281 | void i915_request_module_exit(void) |
2282 | { |
2283 | kmem_cache_destroy(s: slab_execute_cbs); |
2284 | kmem_cache_destroy(s: slab_requests); |
2285 | } |
2286 | |
2287 | int __init i915_request_module_init(void) |
2288 | { |
2289 | slab_requests = |
2290 | kmem_cache_create(name: "i915_request" , |
2291 | size: sizeof(struct i915_request), |
2292 | align: __alignof__(struct i915_request), |
2293 | SLAB_HWCACHE_ALIGN | |
2294 | SLAB_RECLAIM_ACCOUNT | |
2295 | SLAB_TYPESAFE_BY_RCU, |
2296 | ctor: __i915_request_ctor); |
2297 | if (!slab_requests) |
2298 | return -ENOMEM; |
2299 | |
2300 | slab_execute_cbs = KMEM_CACHE(execute_cb, |
2301 | SLAB_HWCACHE_ALIGN | |
2302 | SLAB_RECLAIM_ACCOUNT | |
2303 | SLAB_TYPESAFE_BY_RCU); |
2304 | if (!slab_execute_cbs) |
2305 | goto err_requests; |
2306 | |
2307 | return 0; |
2308 | |
2309 | err_requests: |
2310 | kmem_cache_destroy(s: slab_requests); |
2311 | return -ENOMEM; |
2312 | } |
2313 | |